AI News Roundup: November 12, 2026 - Businesses in ChatGPT, an Anthropic IPO, and a $0.10 Model
A week where the frontier got cheaper, an avatar app got a voice, and the model market started pricing by the job instead of the token.
The week of November 5 to 12 was less about a single big model and more about everyone repositioning around price, agents, and provenance. OpenAI opened 'Businesses in ChatGPT' in beta on November 10, letting companies claim a profile and launch a Business Agent - the clearest sign yet that the product is becoming a storefront, not just a chat. Anthropic launched its IPO roadshow the same day: 825 million shares, an expected $70 to $78 price, and a Nasdaq listing under ANT.
On the model front, Meta released Muse Ember on November 9, its smallest Muse model built for the tool calls and coding subtasks agents repeat at scale, priced at $0.10 per million input tokens. That's the cheapest credible model we've seen aimed at high-volume agent work. Anthropic also shipped Claude Fable 5.2 and Mythos 5.2 the same day - up to 30 percent faster than 5.1, $8 per million input tokens, and a new coding record, with Mythos gated to verified security and life-sciences orgs.
Google pushed Gemini 4 Carbon to paying developers on November 10 at $2 per million input tokens, claiming the top spot on the updated Artificial Analysis Intelligence Index. Gemini Omni 1.2 Flash went generally available on November 5 with 15-second video, conversational editing, and 720p pricing down 30 percent to $0.07 per second. Google also said it will deprecate Gemini 2.5 Pro and 2.5 Flash on December 1, auto-rerouting that traffic to newer models - a reminder that model names are becoming moving targets.
The provenance and open-weights thread advanced too: Thinking Machines Lab released Inkling 2 Small, a 594B-parameter mixture-of-experts model with Apache 2.0 weights promised after testing, and Pika published open weights for H3 Flex, a six-step video-and-audio model post-trained on MiniMax H3. OpenAI's GPT-6 Cyber, released November 5, stays gated to approved defenders, reporting a 48.5 percent ExploitGym score. The pattern: the most capable or most dangerous capabilities are being released to fewer people, not more.
What I'd actually pay attention to: the shift from per-token to per-job pricing. When a model is $0.10 per million tokens, the unit of value stops being the word and becomes the task an agent runs a thousand times. That changes which models make sense for production, not for demos.
Frequently asked questions:
Q: Is Anthropic's IPO confirmed?
A: The roadshow launched November 10 with a stated price range and share count; the listing itself is the next step, not a done deal.
Q: What happens to apps still calling Gemini 2.5 Pro after December 1?
A: Google says those calls will be served and billed by a newer model, so expect different output and a different line item.
Q: Which release matters most for builders?
A: Muse Ember's price point, because it lowers the cost of running agents at volume more than any single benchmark this week.